HPL-AI Version 2.3b

Kan Wu optimized the device blaspp routines through the
strategy of buffering memory: the program will check whether
the buffer size is sufficient at runtime, free and reallocate
the buffer when necessary. Therefore, it is recommended to
run the benchmark twice with the same parameters to obtain
higher computation rate:
-- [M] src/blas/HPLAI_blas.cc
-- [M] testing/ptest/HPLAI_pdinfo.cc